Skip to content

Commit 75f7c45

Browse files
committed
新增解码说明
1 parent 3572aab commit 75f7c45

File tree

1 file changed

+21
-0
lines changed

1 file changed

+21
-0
lines changed

docs/model_zoo.md

Lines changed: 21 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -208,4 +208,25 @@ yolov5采用dataloader进行测试时候,实际上是有自适应的,虽然
208208
可以看出yolov5这种设置会更好一点。应该就是这个差异导致的mAP不一样。
209209

210210

211+
关于yolov5中心点解码问题的分析:
212+
yolov5中是:
213+
x_center_pred = (pred_bboxes[..., 0] * 2. - 0.5 + grid[:, 0]) * stride # xy
214+
y_center_pred = (pred_bboxes[..., 1] * 2. - 0.5 + grid[:, 1]) * stride # xy
211215

216+
pred_bboxes的xy是相对网格左上角的偏移,而mmdetection中anchor是相对网格中心,以x方向为例
217+
(x*2-0.5+grid_x)*s
218+
= (x*2-0.5)*s+grid_x*s # yolov5
219+
= (x*2-0.5)*s+(xcenter/s-0.5)*s
220+
= (x-0.5)*2*s+xcenter # mmdetection
221+
222+
223+
关于yolov4的scale_xy,中心点解码问题的分析:
224+
yolov4中是:
225+
x_center_pred = (pred_bboxes[..., 0] * self.scale_x_y - 0.5 * (self.scale_x_y - 1) + grid[:, 0]) * stride # xy
226+
227+
pred_bboxes的xy是相对网格左上角的偏移,而mmdetection中anchor是相对网格中心,以x方向为例
228+
229+
(x*1.05-0.5(1.05-1)+grid)*s
230+
= (x*1.05-0.5(1.05-1))*s+grid_x*s
231+
= (x*1.05-0.5(1.05-1))*s+(xcenter/s-0.5)*s
232+
= (x*1.05-0.5(1.05-1)-0.5)*s+xcenter

0 commit comments

Comments
 (0)